th

The 14 World Conference on Earthquake Engineering

October 12-17, 2008, Beijing, China

MODELI&G PROGRESSIVE COLLAPSE I& REI&FORCED CO&CRETE

FRAMED STRUCTURES
Khalid M. Mosalam1, Mohamed Talaat2, and Sangjoon Park3
1

Professor and Vice Chair, Department of Civil and Environmental Engineering, University of California,
Berkeley, CA 94720-1710, U.S.A., Email: mosalam@ce.berkeley.edu
2
Senior Engineer, Simpson Gumpertz and Heger Inc., San Francisco, U.S.A., mtalaat@sgh.com
3
Graduate Student, Dept. of Civil & Env. Eng., Univ. of California, Berkeley, U.S.A., sangjoon@berkeley.edu
ABSTRACT: The paper presents an analytical formulation of an element removal algorithm based on dynamic
equilibrium and the resulting transient change in system kinematics. The algorithm is implemented into an
open-source finite element code, is numerically validated using a benchmark problem with simple element
removal criteria, and is able to capture the effect of uncertainty in member capacity. Realistic element removal
criteria are introduced for mode-dependent gravity load collapse of seismically deficient and retrofitted
reinforced concrete (RC) columns. An application is presented using a RC 3-bay, 3-story frame previously
tested on a shaking table. In this application, comparison between the computationally predicted and
experimentally observed collapse modes is presented giving confidence in the developed analytical formulation.

KEYWORDS: Collapse, Dynamic analysis, Element removal, Frame, Reinforced concrete

1. I&TRODUCTIO&
Progressive collapse assessment using nonlinear time-history finite element (FE) simulation is gaining popularity
over traditional methods based on alternate path analysis and redundancy detailing. Review of applications of
these traditional methods can be found in [1]. There is limited literature on this developing field of research as few
experimental investigations were conducted on reduced-scale reinforced concrete (RC) frames that are redundant
enough to experience progressive collapse, with seismically deficient RC columns designed to lose axial load
capacity, e.g. [2,3]. Experimental data from these studies are intended to establish the behavior of RC systems
after the loss of one or more columns. However, observed collapse modes are limited and difficult to generalize to
cases involving disconnection and possible subsequent collision by collapsed elements.
Several analytical approaches were developed for modeling progressive collapse [1]. To capture the progression
of collapse, one must model the behavior of components under extreme loading and identify the criteria for their
removal. Two modes of RC column collapse, namely shear-axial and flexure-axial, are presented. Shear-axial
collapse was investigated in several experimental and analytical studies [3]. The shear friction-based model in [4]
is utilized to establish limit-states for shear-axial collapse. The damage indices in [1] for fiber-discretized
cross-sections are used for flexure-axial collapse. For non-ductile and retrofitted using fiber-reinforced polymers
(FRP) columns, new models are adopted from [1] for confined fiber-discretized cross-section and
confinement-sensitive concrete material, and from [1] for confinement-sensitive bar buckling and lap splices. The
computational implementation was carried out in OpenSees [5]. Finally, the RC 3-bay, 3-story frame from [3] is
adopted as an application of the developed computational approach for collapse modeling with element removal.

2. DIRECT ELEME&T REMOVAL: MECHA&ICS, IMPLEME&TATIO&, A&D VALIDATIO&

Consider the dynamic equilibrium of the node in Figure 1a. At time t , beams B1 & B 2 and columns C1 & C 2
are attached to the node. The free-body diagram is shown without externally applied or damping nodal forces. F
and M refer respectively to resisting forces (both shear and axial) and bending moments from the attached

th

The 14 World Conference on Earthquake Engineering

October 12-17, 2008, Beijing, China

elements. The node has a lumped translational mass m and a rotational mass moment of inertia I . Dynamic
equilibrium is satisfied by the nodal inertia forces ma x , ma y , and I , where a x and a y are translational
accelerations in the respective x and y directions and is rotational acceleration in the x y plane.
MC2,FC2

MC2,FC2

Mass = m
Time = t

MB1,FB1

Mass = m
Time = t = t+t

MB2,FB2

MB1,FB1

MB2,FB2

max

X
X

max

may

may

MC1,FC1

ax >> ax
ay >> ay
>>

Removed element
releases internal forces

(a) Before element removal

(b) After element removal
Figure 1 Dynamic equilibrium of a node connected to a collapsed element
At time t , column C1 reaches its collapse limit-state and is removed from the computational model at time
t = t + t where t is the time step. The resulting free body diagram is shown in Figure 1b, where the nodal
masses and accelerations are updated to m , I , a x , a y , and , respectively. Practically, the majority of the
lumped nodal masses are derived from the floor system, and the updated masses may not be significantly different
from the original ones. The dynamic equilibrium of the node under resisting and inertia forces is expressed by

Fx max
Fx max
v

v

Pex + Fy ma y = Pex + Fy may

M I
M I

(1)

v
where Pex is vector of externally applied nodal loads, independent of the element connectivity and not affected
by the removal of a collapsed element. The summation is for the elements connected at the node. Upon abrupt
element collapse and removal, this dynamic equilibrium is disturbed by the sudden release of internal forces from
C1 and must be restored before the solution continues. The resisting forces in the connected elements, F & M ,
can only change as a result of updated element deformations, which require changes in relative displacements
(and possibly velocities) between the elements end-nodes. These changes result from updated nodal velocities
and accelerations and do not take place instantaneously. The inertia forces are directly related to nodal
accelerations a x , a y , and , which need to change abruptly to satisfy dynamic equilibrium at time t . Since
the connected elements end-nodes other than their shared node are under dynamic equilibrium themselves, this
abrupt change is first localized at the shared node at time t . In subsequent time steps, element removal
superposes transient loading on the damaged system, whereby the resulting changes in nodal kinetics lead to an
updated set of resisting (and inertia) forces propagating outward through the connected elements into neighboring
nodes. These updated nodal forces must satisfy dynamic equilibrium of their respective degrees of freedom
(DOFs) during every time step resulting in updated inertia forces and nodal accelerations. This process continues
to propagate in the damaged system until updated nodal configurations for a new equilibrium state are reached.
Upon reaching the new equilibrium state, the nodal velocities of the damaged system result in displacement
overshooting, amplifying deformation demand in the structural elements. In the hypothetical absence of external
excitation, free vibration of the system ensues. An initial transient phase is followed by a steady-state phase
oscillating about the new equilibrium state. The new deformation demands may lead to collapse of other elements,

th

The 14 World Conference on Earthquake Engineering

October 12-17, 2008, Beijing, China

which would be safe in the new equilibrium state using quasi-static analysis. If so, a second transient phase is
excited, and this process continues until the system either (a) reaches an equilibrium state about which it can
safely oscillate and collapse progression is arrested; (b) undergoes overall gravity load collapse; or (c) undergoes
partial collapse which is nevertheless compartmentalized while the rest of it survives. In case (c), the structure will
probably need to be ultimately demolished, yet prospect of maintaining life safety is enhanced compared to case
(b). In the presence of an earthquake excitation, the input ground motion is treated as equivalent external loads.
The input time step may be larger than adequate for transient effects, particularly if reaching a new equilibrium
state requires local axial vibrations in stiff structural elements with significantly short natural periods. Hence,
shorter time steps are needed immediately after element removal, during which the input ground motion is
interpolated. An adaptive time stepping scheme can be used for computational efficiency [1].
An algorithm is designed and implemented for automated removal of collapsed elements during an ongoing
simulation, Figure 2. The implementation is carried out as a new OpenSees module, designed so that it is called by
the main analysis module after each converged load step, and ordered to check each element for possible violation
of its removal criteria. A violation of any criterion triggers the activation of the element removal algorithm before
returning to the main analysis module. The algorithm is split into two main sets of procedures. The first set is fully
implemented within the source-code of OpenSees. It includes updating nodal masses, checking if the removal of
the collapsed element results in leaving behind dangling nodes or floating elements, Figure 3, which must be
removed, and removing all associated element and nodal forces, imposed displacements, and constraints. The
second set involves tracking motion of removed element for possible collision. The kinetics of the separated
end-node(s) at the time of separation are identified from the last converged state of the system, representing the
initial conditions for free falling or rotating motion of the collapsed element. This motion is tracked as a rigid body
motion of the element under effects of its own weight, and compared to the position of the damaged system to
determine if and when collision takes place. At the time of collision, the relative impact velocity is computed and
used to determine the impact duration and the magnitude and temporal variation of impact force to be superposed
on the damaged system during subsequent time steps. This second set of procedures is performed in the present
study by conducting the necessary computations using MATLAB [6], and inputting the results to OpenSees.

Start
from
main
code
Italicized text
executed outside
of OpenSees

Check for dangling nodes,

floating elements, and
element loads and masses

Remove dangling nodes

Remove floating elements
Delete element/node loads
Remove element
Update nodal masses

Identify nodal kinetics

at separation

Update structural model, time

step, and solution parameters

Track collapsed element

motion until impact

Identify location, compute force,

duration, and mass redistribution

End
back to
main
code

Figure 2 Automatic element removal algorithm

After completing both sets of procedures, the algorithm updates the system state to the new masses, geometry and
impact forces. The assumption of Rayleigh viscous damping (mass and/or stiffness proportional) is implicitly
affected. Since numerical convergence may face difficulties following extreme events, especially in softening
structural systems, the solution parameters may need to be updated. This is conducted by iteratively switching
solver type, convergence criteria, and other analysis options, e.g. time step, so that an ultimate failure to converge
would likely take place only if the damaged structural system experiences global instability. An OpenSees input
script for adaptive time stepping and solver modification is developed and used in the examples presented herein.
The above algorithm implementation is numerically validated using a benchmark problem representing an
idealized structural system consisting of two trussed cantilever beam canopies, Figure 4, with properties (elastic
modulus, E , cross-sectional area, A , and moment of inertia, I ) in Table 1. The truss members are linear elastic
up to axial load Fy , then become either brittle or perfectly plastic up to a ductility limit, Table 2, before removal.
Four cases are used to explore sensitivity of the algorithm to uncertainty in element capacities. The beam elements
are linear elastic. Nonlinear geometry and large deformation effects are included in the FE model by co-rotational
geometric transformation [5]. Masses are lumped as 10 ton at node 1, 6 ton at node 8, and 4 ton at nodes 6, 7, and

th

The 14 World Conference on Earthquake Engineering

October 12-17, 2008, Beijing, China

9-11. Gravity loads are applied followed by excitation at nodes 2-5 and 12 in the vertical direction using the
earthquake record from Northridge 1994, Tarzana, 90 component, with scaled peak ground acceleration of 1.04g.
A time step of 0.005 sec is chosen with no viscous damping. Collapsed truss elements separate from the structural
system at the connection with the beams. Hence, collapsed element 3 falls and collides with the lower canopy at
node 9. After collision, element 3 deforms in flexure as simply supported at nodes 4 and 9. The energy of the
impact dissipates during one half-cycle of flexural deformation before rapidly decaying, e.g. due to friction, and
element 3 remains supported by nodes 4 and 9 until collapse of the lower canopy. Collision with (or loss of) a truss
element at any node results in updating this nodes mass and gravity load by adding (or subtracting) 2 tons.
10
3
1
10
2
2

1
10
3

Intact structure

Nodal loads

6
6

5
4
Dangling
node
Floating
element

7
8

8 9
9 10 11

5
10
11

Dashes represent
collapsed elements

12

Figure 3 Elements, nodes, and loads requiring

removal due to element removal

18

610/3=20

Figure 4 Geometry (in meters) and FE model of

benchmark problem

Table 1 Element properties of benchmark problem

Element EA [kN] EI [kNm2] Fy [kN]
1
2
3
4-9
10,11

2,000
100,000
1,000
100,000
3,000

3,000
5,000
1,000
10,000
5,000

100
(Elastic)
50
(Elastic)
150

Table 2 Ultimate truss element ductility limits for benchmark problem

Truss element Case 1 Case 2 Case 3 Case 4
1, 11
3
10

5
5
5

5
5
1

5
1
5

5
1
1

All simulations were successfully conducted without exhibiting numerical convergence problems. Cases 1 and 2
result in similar responses where no element collapse takes place. In both cases, the upper canopy elements 1 and
3 exhibit brief yielding, inducing a ductility demand of 1.17 and 1.11, respectively. Cases 3 and 4 result in
collapse of element 3 onto the lower canopy at 8.875 sec. As a result, the ductility demand on element 1 increases
to 2.1 and the upper canopy exhibits no further collapse. The collapse of element 3 results in larger permanent
displacement in the vertical direction at node 1, Figure 5a where nodal displacements are normalized using
respective values caused by the gravity load. In the lower canopy, element 3 collides with node 9 at 11.26 sec, and
the resulting demand causes yielding in elements 10 and 11. In Case 3, the ductility capacity in both elements is
sufficient to prevent the progression of collapse in the system, which reaches a new equilibrium state, Figure 5b.
In Case 4, brittle element 10 collapses at 12.37 sec. Shortly afterwards, element 11 reaches its ultimate ductility
limit and collapses at 14.57 sec leading to complete collapse of the lower canopy, Figure 5b. The deformed shapes
are illustrated in Figure 6. There is clear contrast between the small difference in ductility demands for elements 3

th

The 14 World Conference on Earthquake Engineering

October 12-17, 2008, Beijing, China

and 10 and the large difference in simulated progressive collapse outcome. This highlights the sensitivity of the
simulation procedure, due to path dependency, to uncertainty in the structural model parameters. Hence,
simulation-based progressive collapse assessment needs to be conducted within a probabilistic framework [1].
-0.5

Time = 8.875

Normalized Displacement

Normalized Displacement

-0.5

-1

-1.5

Cases 3&4

-2.5
0

10
15
Time [sec]

20

-1

-1.5

Cases 1&2

-2

Time = 11.260

Cases 1&2

-2

-2.5
0

25

Case 3
Case 4

10
15
Time [sec]

20

25

(a) Node 1
(b) Node 11
Figure 5 Simulated nodal displacement time-histories for benchmark problem

10

15

20

25

10

15

20

25

10

15

20

25

(a) Cases 1&2

(b) Case 3
(c) Case 4
Figure 6 Snapshots of benchmark structural model deformation (no amplification) at Time=14.5 sec

3. ELEME&T REMOVAL CRITERIA

Three element removal criteria are defined for two different modes of failure in seismically deficient and
retrofitted RC beam-columns and for truss members. These criteria are implemented in OpenSees for force-based
and displacement-based distributed plasticity fiber elements, lumped plasticity beam-column elements with
fiber-discretized plastic hinges, shear-axial coupled spring elements [4] and truss elements.
A set of material damage indices is developed and calibrated in [1] for fibers presenting reinforcing steel bars,
lap-spliced bars, and concrete (confined and unconfined). These damage indices are based, respectively, on the
accumulation of steel plastic strains, the maximum bond-slip displacement, and the ratio of hysteretic energy
dissipation to compressive fracture energy. These material-level damage indices define the following two
aggregated cross-section damage indices to reflect respective loss of axial load and bending moment capacities.

D A = 1 conf 1 (D fiber Afiber A cross - section ) ,

fibers

2
DM = 1 conf 1 Dfiber Afiber hfiber
I cross-section
fibers

(2a,b)

where A and I refer to transformed area and moment of inertia, and h to the distance between the fibers
center and the uncracked section centroid. 0 Dfiber 1 is the material damage index defined for individual
fibers [1]. conf is a loss of confinement indicator with 0.0 value if confining medium (e.g. transverse reinforcing
ties or FRP jacket) fracture is detected and 1.0 otherwise. These damage indices are suitable for use with deficient
and FRP-retrofitted RC columns. When DA = 1 , the associated column cross-section is considered to have lost its
axial load capacity and the element removal algorithm is invoked. A future extension may explicitly reflect the
axial load level in the column. An analogous approach can be defined for DM , but is not of interest in this paper.

th

The 14 World Conference on Earthquake Engineering

October 12-17, 2008, Beijing, China

An element removal criterion is developed to correspond to violating limit equations for the drift capacity model
in [4] where shear-axial coupled springs are connected in series to a beam-column element to construct a
compound element. When the axial limit curve is reached in a RC column, the corresponding spring is removed.
Consequently, the connectivity between the associated beam-column element and the structural system is severed
at one end-node. Similarly, when the shear limit curve is reached in a RC beam, the corresponding spring is
removed. Hence, a shear-damaged RC column is allowed to maintain its axial load while shedding shear force
until it reaches the axial load-lateral drift limit-state, while a shear-damaged RC beam is assumed to collapse in a
brittle manner due to absence of a compressive axial force that may prevent shear failure from rapid propagation.
A simple element removal criterion is developed for violating maximum (positive) or minimum (negative) axial
deformation values. When the axial deformation in a truss element reaches either threshold value defined by the
analyst, the element is removed. The choice of such values enables modeling both brittle and ductile behavior [1].

4. RC 3-BAY, 3-STORY FRAME APPLICATIO&

The structural system is selected to take advantage of existing data from a shaking table test conducted at UC
Berkeley. This 2D RC 3-bay, 3-story frame structure was constructed to examine the behavior of a non-ductile
RC office building typical of the 1960s and 1970s designs in California. The test frame was 1/3-scale of a
prototype structure, Figure 7a. Two of the columns in the frame had non-ductile details with widely spaced ties,
and 90 hooks, while the other two columns followed ACI318-2002 requirements [7]. Beam designs were based
on the weak-column strong-beam mechanism using beams with increased depth. Thus, joint shear stresses were
effectively decreased. The dimension and reinforcement of each column were identical along all stories and
those of beams were the same over all floors and spans. Summaries of the geometrical and material properties
are given in Tables 3 and 4, respectively. Complete details of the frame design and test results are given in [3].
A corresponding FE model, Figure 7b, is developed and used to perform the collapse analysis. The main objective
of this analysis is investigating the collapse mode, causes, and sequence. Displacement-based beam-column
elements for the RC frame and zero length elements of coupled shear-axial limit state material [4] at the bases of
the two columns to the east in all stories comprised the FE model. In Figure 7b, springs are shown only at the
foundation level for clarity. Meshing, fiber-discretization, and constitutive modeling are similar to those of
applications A and B in [1]. Nodes 1-4 are fully restrained and masses and gravity loads are lumped at the nodes.
Co-rotational geometric transformation is used for all elements. The accuracy of the FE model is verified by
comparing the first three periods of vibration from the analysis to those from the snap-back tests [3]. From Table 5,
the maximum difference of 15% is mainly due to differences in real concrete material stiffness and that used in the
simulation. Mass- and stiffness-proportional viscous damping (=0.05) is assumed for the first and third modes.

D1
1

Z2

Y2

Z1

Y1

X2

X1

Coupled
shear-axial
B1
A1
limit state springs

C1
2

(a) 1/3 scale Test frame [3]

(b) OpenSees modeling
Figure 7 Shaking table test frame and OpenSees FE model
For the purpose of this application, the ground motion recorded from the frame footings during the first run of
the shaking table test is used as the reference ground motion in the analysis. This motion was the Chile

th

The 14 World Conference on Earthquake Engineering

October 12-17, 2008, Beijing, China

earthquake record at Valparaiso, 1985, scaled up by a factor 4.06. To induce collapse of columns A1 and B1,
this scaled up ground motion needed to be scaled further by a factor of 1.43, i.e. total scale of 4.061.43=5.8, in
the third run of the shaking table test [3]. It is to be noted that the shaking table test was performed in three runs
and the third (collapse level) run took place on the damaged state from the first two runs without any repair.
Therefore, a higher additional scale factor of 2.0, i.e. total scale of 4.062=8.12, is needed in the computational
study to cause the collapse mechanism of the frame as the initial damage was not incorporated in this analysis.
Table 3 Geometry and details of shaking table test structure [3]
Structural member
Dimensions [mm]
Ductile
152152
Column
Non-ductile
152152
Beam

152229

Main reinforcement
Eight 10-mm (#3) bars

Closed ties or stirrups

3.2-mm wire@32mm overall height

Eight 6-mm (#2) bars

3.2-mm wire@100mm overall height

Four 10-mm (#3) bars top & bottom

3.2-mm wire@89mm overall span

Table 4 Material properties of shaking table test structure [3]

Material
property

Concrete
f c [MPa]

Concrete
f t [Mpa]

Concrete
cover[mm]

Value

25

2.7

17

f y [Mpa]

Steel bars yield stress

10-mm
441

6-mm
483

3.2-mm
655

Es, average [Mpa]

1.85105

Table 5 Comparison between analysis and snap-back test results of the shaking table test structure
Period
Analysis
Snap-back test
Ratio

T1 [sec]
0.348
0.303
1.15

T2 [sec]
0.113
0.101
1.12

T3 [sec]
0.069
0.069
1.00

.
The simulated system response eventually exhibits global collapse due to soft first story mechanism after axial
failure of the two non-ductile columns. In Figure 8, the story drift ratio is the average drift ratios from the four
pairs of nodes defining story columns. Columns A1 and B1 are subjected to shear failure in the interval 7.5-8.0
sec yet are still carrying axial loads. At 17.14 sec, shear-damaged column A1 loses its axial load capacity and is
removed, closely followed by B1 removal, and the first story drift ratio disproportionately increases compared
to the third story. Subsequently, the first story beams act as a cantilever and bending moment is localized at the
top of column C1 resulting in beam Y1 collapse due to loss of confinement [1] and shortly followed by the
second story beam Y2 collapse. Finally, the frame completely collapses after removing the last two columns in
the first story, i.e. C1 and D1. The time at incipient collapse is the time when unrestrained motion is initiated at
17.165 sec (removal of B1), and the maximum meaningful first story, third story, and overall drift ratios prior
to this incipient collapse are 0.0472, 0.0120, and 0.0425, respectively. The mode and sequence of collapse are
given in Table 7. Figure 9 compares the analytical and experimental collapse modes at their inceptions.
0.2
0.15

1st story

0.1

3rd story
Overall

Remove
Column A1
Column B1

Drift ratio

0.05
0
-0.05
-0.1
-0.15

Collapse

-0.2
-0.25

8
10
Time(sec)

12

14

16

18

Figure 8 Drift ratio time histories obtained from the OpenSees FE model

th

The 14 World Conference on Earthquake Engineering

October 12-17, 2008, Beijing, China

Table 6 Comparison of computationally simulated and experimentally observed sequence of collapse

Computation
Experiment
Collapse mode
Shear-axial
Flexural (confinement) loss
Shear-axial
Collapsed FE
A1
B1
Y1
Y2
C1
D1
B1
A1
Time [sec]
17.1400 17.1650 17.6160 17.9732
18.1132 18.3794 Run 1 Run 2

Scale = 4.06 causing shear failure but loss of gravity load capacity took place in run 3 (scale = 5.8) [3]

Remove
Column A1
Column B1

(a) Test collapse mode [3]

(b) Analysis collapse mode
Figure 9 Comparison of collapse mode

7. CO&CLUDI&G REMARKS
An analytical formulation is developed based on the dynamic equilibrium of forces to account for abrupt removal
of collapsed elements from a structural model. The algorithm simulates the dynamic redistribution of forces in
addition to a simplified modeling of impact. It is implemented and numerically validated using a benchmark
problem. Element removal criteria are introduced for mode-dependent axial load collapse in reinforced concrete
columns and for truss members. An application is presented using a RC 3-bay, 3-story frame with light transverse
reinforcement. Comparisons between the computationally estimated and experimentally observed from previous
shaking table tests for the collapse modes and results revealed the validity of the developed computational
approach of progressive collapse simulation using direct element removal.

8. ACK&OWLEDGEME&TS
This study was supported by the Earthquake Engineering Research Centres Program of the NSF under award
#EEC-9701568 and by the NSF award #0618804 through the Pacific Earthquake Engineering Research Center
(PEER). Opinions and findings are those of the author(s) and do not necessarily reflect those of the NSF.
REFERE&CES
1. Talaat M. and Mosalam KM. (2008). Computational modeling of progressive collapse in reinforced concrete
frame structures. PEER Technical Report 2007/10.
2. Kim Y. and Kabeyasawa T. (2004). Dynamic test and analysis of an eccentric reinforced concrete frame to
collapse. 13th World Conference on Earthquake Engineering, Vancouver, BC, Canada.
3. Ghannoum W. (2007). Experimental and analytical dynamic collapse study of a reinforced concrete frame
with light transverse reinforcements. PhD Dissertation, University of California, Berkeley.
4. Elwood K.J. and Moehle J.P. (2005). Axial capacity model for shear-damaged columns. ACI Structural
Journal, 102:4, 578-587.
5. Mazzoni S., McKenna F., Scott, M. and Fenves, G. (2006). OpenSees User Command-Language Manual.
http://opensees.berkeley.edu/OpenSees/manuals/usermanual/index.html.
6. The MathWorks, Inc. (2007). Getting Started with Matlab 7 (Release 2007b). Natick, MA.
7. ACI. (2002). Standard building code requirements for reinforced concrete. ACI 318-02. Detroit, MI.